They Are Stealing or Buying Our Info in Any Way Possible: Fitness classes for robots --- on the path to the universal machine scientists are teaching robots human movements in specially designed training halls.

“Residents of New York can now take advantage of a unique service: using an app called Shift to order professional cleaners to tidy their apartments—free of charge.

 

The company behind Shift, the Aachen-based startup MicroAGI, assures users there are no hidden costs—and the dirtier the apartment, the better. The only condition is that you must allow the cleaners to wear a camera on their foreheads while working to record every movement. After all, the company is less interested in the cleaning itself; it is seeking training data for robots. On another Shift website, people can sign up to film themselves—specifically their hands and their actions—whether at home or at work. It is probably best not to dwell on data privacy issues in this context.

 

Following the success of large language models, there is now much talk of a "ChatGPT moment" for robotics. If robots had access to chatbots capable of answering—with varying degrees of reliability—any question about how to solve a problem, couldn't these artificial helpers finally move into factories and households? It is not quite that simple. Language models are exactly what their name implies: models of language. They are not models of actions; they provide information but not control commands. Nevertheless, they are useful: they enable communication with machines and allow robots to identify objects—for instance, figuring out on their own that a banana peel found on the floor belongs in the trash. Language models help robots break tasks down into sub-tasks and plan their execution—such as placing a glass down before pouring the wine. However, they are insufficient for navigating robots through the real world.

 

To change this, intensive work is being done on foundation models conducting research for robotics. Data is the bottleneck; there is a shortage of roughly 100,000 years' worth of training data.

 

While the internet is teeming with videos, that doesn't really help. "You can derive the kinematics of movements from videos, but not the dynamics," explains Lorenzo Masia, Professor of Intelligent Bio-Robotic Systems at the Technical University of Munich (TUM). In other words, while films show what a grasping motion looks like, they don't reveal how firmly one must hold a banana peel so that it neither drops nor gets squashed.

 

One way to obtain such data is through simulation. A computer model of the robot is created, and it is tasked with learning the required actions. This process is faster and less expensive than working with actual robots. However, a gap remains between simulation and reality—the "sim-to-real gap." "Doing this in a simulation will never be as precise as the data obtained from actual movements," says Masia.

 

MicroAGI relies on data generated by filming from the perspective of the person performing the action.

 

Others collect data while teleoperating robots; humans control the robot using devices like data gloves, performing grasping motions with their hands that the robotic hand then replicates. The movements and the forces applied are recorded. The resulting data can then be used to train other robots. The aim is to create a kind of database of actions—a collective memory that a robot facing a specific task can access to select the appropriate action. At the robot manufacturer NEURA Robotics, this data collection is known as the "Neuraverse."

 

To accelerate data generation and training, NEURA Robotics is developing training facilities for robots, known as "Robogyms." The world's largest facility of this kind—a "gym" for machines—is currently being built in collaboration with the Technical University of Munich in Convergence Center at Munich Airport. Spanning over 2,000 square meters, the facility is setting up various tasks that robots are to learn to perform. "We opted for the 'gym' concept because there is an urgent need for data—data for foundation models and for industrial applications," explains Lorenzo Masia, who will head the facility.

 

"Suppose a company wants to automate part of a production line. Customers could come to us; we would replicate the task, train the robots via teleoperation, and collect the data. Finally, we would install the system—that is the idea behind the Robogym." If customers are willing to share their data with others, the cost of the "robot fitness course" is reduced.

 

Ideally, the data collected in this way could eventually be used to train all kinds of robots—from manufacturing arms to quadrupedal or humanoid robots. The concept behind this "multi-embodiment learning" is that the less a foundation model is tied to a specific robot body, the more versatile it becomes.

 

According to one vision, humanoid robots could eventually become true general-purpose machines—capable of working in any job designed for humans, moving from the assembly line to the warehouse, and even bringing back coffee for the remaining human workers on their way back. This would also bring the much-discussed domestic robot within the realm of possibility. Such general-purpose machines could also represent a step toward the long-sought goal of Artificial General Intelligence—an intelligence capable of handling any task.

 

Others, however—including Yann LeCun, Meta’s former chief scientist—dismiss the very idea of ​​such general intelligence as nonsense, whether in humans or machines. Humans are convinced they possess such universal intelligence only because they overlook their own limitations. Instead of aiming for a general intelligence—ideally one that is superhuman—we should focus on superhuman adaptability and on specialized components that work together as effectively as possible. The systems that fold our proteins will not be the same ones that fold our laundry, the researchers argue.

 

The future of robots, therefore, depends not only on the availability of training data but also on the underlying definition of intelligence. Once again, it is artificial intelligence that compels us to think more deeply about natural intelligence.” [1]

 

They Are Stealing or Buying Our Info in Any Way Possible: Fitness classes for robots ---  on the path to the universal machine scientists are teaching robots human movements in specially designed training halls.

 

“Residents of New York can now take advantage of a unique service: using an app called Shift to order professional cleaners to tidy their apartments—free of charge.

 

The company behind Shift, the Aachen-based startup MicroAGI, assures users there are no hidden costs—and the dirtier the apartment, the better. The only condition is that you must allow the cleaners to wear a camera on their foreheads while working to record every movement. After all, the company is less interested in the cleaning itself; it is seeking training data for robots. On another Shift website, people can sign up to film themselves—specifically their hands and their actions—whether at home or at work. It is probably best not to dwell on data privacy issues in this context.

 

Following the success of large language models, there is now much talk of a "ChatGPT moment" for robotics. If robots had access to chatbots capable of answering—with varying degrees of reliability—any question about how to solve a problem, couldn't these artificial helpers finally move into factories and households? It is not quite that simple. Language models are exactly what their name implies: models of language. They are not models of actions; they provide information but not control commands. Nevertheless, they are useful: they enable communication with machines and allow robots to identify objects—for instance, figuring out on their own that a banana peel found on the floor belongs in the trash. Language models help robots break tasks down into sub-tasks and plan their execution—such as placing a glass down before pouring the wine. However, they are insufficient for navigating robots through the real world.

 

To change this, intensive work is being done on foundation models conducting research for robotics. Data is the bottleneck; there is a shortage of roughly 100,000 years' worth of training data.

 

While the internet is teeming with videos, that doesn't really help. "You can derive the kinematics of movements from videos, but not the dynamics," explains Lorenzo Masia, Professor of Intelligent Bio-Robotic Systems at the Technical University of Munich (TUM). In other words, while films show what a grasping motion looks like, they don't reveal how firmly one must hold a banana peel so that it neither drops nor gets squashed.

 

One way to obtain such data is through simulation. A computer model of the robot is created, and it is tasked with learning the required actions. This process is faster and less expensive than working with actual robots. However, a gap remains between simulation and reality—the "sim-to-real gap." "Doing this in a simulation will never be as precise as the data obtained from actual movements," says Masia.

 

MicroAGI relies on data generated by filming from the perspective of the person performing the action.

 

Others collect data while teleoperating robots; humans control the robot using devices like data gloves, performing grasping motions with their hands that the robotic hand then replicates. The movements and the forces applied are recorded. The resulting data can then be used to train other robots. The aim is to create a kind of database of actions—a collective memory that a robot facing a specific task can access to select the appropriate action. At the robot manufacturer NEURA Robotics, this data collection is known as the "Neuraverse."

 

To accelerate data generation and training, NEURA Robotics is developing training facilities for robots, known as "Robogyms." The world's largest facility of this kind—a "gym" for machines—is currently being built in collaboration with the Technical University of Munich in Convergence Center at Munich Airport. Spanning over 2,000 square meters, the facility is setting up various tasks that robots are to learn to perform. "We opted for the 'gym' concept because there is an urgent need for data—data for foundation models and for industrial applications," explains Lorenzo Masia, who will head the facility.

 

"Suppose a company wants to automate part of a production line. Customers could come to us; we would replicate the task, train the robots via teleoperation, and collect the data. Finally, we would install the system—that is the idea behind the Robogym." If customers are willing to share their data with others, the cost of the "robot fitness course" is reduced.

 

Ideally, the data collected in this way could eventually be used to train all kinds of robots—from manufacturing arms to quadrupedal or humanoid robots. The concept behind this "multi-embodiment learning" is that the less a foundation model is tied to a specific robot body, the more versatile it becomes.

 

According to one vision, humanoid robots could eventually become true general-purpose machines—capable of working in any job designed for humans, moving from the assembly line to the warehouse, and even bringing back coffee for the remaining human workers on their way back. This would also bring the much-discussed domestic robot within the realm of possibility. Such general-purpose machines could also represent a step toward the long-sought goal of Artificial General Intelligence—an intelligence capable of handling any task.

 

Others, however—including Yann LeCun, Meta’s former chief scientist—dismiss the very idea of such general intelligence as nonsense, whether in humans or machines. Humans are convinced they possess such universal intelligence only because they overlook their own limitations. Instead of aiming for a general intelligence—ideally one that is superhuman—we should focus on superhuman adaptability and on specialized components that work together as effectively as possible. The systems that fold our proteins will not be the same ones that fold our laundry, the researchers argue.

 

The future of robots, therefore, depends not only on the availability of training data but also on the underlying definition of intelligence. Once again, it is artificial intelligence that compels us to think more deeply about natural intelligence.” [1]

 

1. Fitnesskurse für Roboter: Auf dem Weg zur Universalmaschine: Wissenschaftler bringen Robotern in eigens geschaffenen Trainingshallen menschliche Bewegungen bei. Frankfurter Allgemeine Zeitung; Frankfurt. 01 July 2026: N4.        MANUELA LENZEN